added 2.6.29.6 aldebaran kernel
[nao-ulib.git] / kernel / 2.6.29.6-aldebaran-rt / drivers / infiniband / hw / mlx4 / cq.c
blobde5263beab4a721a8c261b8c8403996bb54e26b8
1 /*
2 * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
13 * conditions are met:
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
17 * disclaimer.
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
34 #include <linux/mlx4/cq.h>
35 #include <linux/mlx4/qp.h>
37 #include "mlx4_ib.h"
38 #include "user.h"
40 static void mlx4_ib_cq_comp(struct mlx4_cq *cq)
42 struct ib_cq *ibcq = &to_mibcq(cq)->ibcq;
43 ibcq->comp_handler(ibcq, ibcq->cq_context);
46 static void mlx4_ib_cq_event(struct mlx4_cq *cq, enum mlx4_event type)
48 struct ib_event event;
49 struct ib_cq *ibcq;
51 if (type != MLX4_EVENT_TYPE_CQ_ERROR) {
52 printk(KERN_WARNING "mlx4_ib: Unexpected event type %d "
53 "on CQ %06x\n", type, cq->cqn);
54 return;
57 ibcq = &to_mibcq(cq)->ibcq;
58 if (ibcq->event_handler) {
59 event.device = ibcq->device;
60 event.event = IB_EVENT_CQ_ERR;
61 event.element.cq = ibcq;
62 ibcq->event_handler(&event, ibcq->cq_context);
66 static void *get_cqe_from_buf(struct mlx4_ib_cq_buf *buf, int n)
68 return mlx4_buf_offset(&buf->buf, n * sizeof (struct mlx4_cqe));
71 static void *get_cqe(struct mlx4_ib_cq *cq, int n)
73 return get_cqe_from_buf(&cq->buf, n);
76 static void *get_sw_cqe(struct mlx4_ib_cq *cq, int n)
78 struct mlx4_cqe *cqe = get_cqe(cq, n & cq->ibcq.cqe);
80 return (!!(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
81 !!(n & (cq->ibcq.cqe + 1))) ? NULL : cqe;
84 static struct mlx4_cqe *next_cqe_sw(struct mlx4_ib_cq *cq)
86 return get_sw_cqe(cq, cq->mcq.cons_index);
89 int mlx4_ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
91 struct mlx4_ib_cq *mcq = to_mcq(cq);
92 struct mlx4_ib_dev *dev = to_mdev(cq->device);
94 return mlx4_cq_modify(dev->dev, &mcq->mcq, cq_count, cq_period);
97 static int mlx4_ib_alloc_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int nent)
99 int err;
101 err = mlx4_buf_alloc(dev->dev, nent * sizeof(struct mlx4_cqe),
102 PAGE_SIZE * 2, &buf->buf);
104 if (err)
105 goto out;
107 err = mlx4_mtt_init(dev->dev, buf->buf.npages, buf->buf.page_shift,
108 &buf->mtt);
109 if (err)
110 goto err_buf;
112 err = mlx4_buf_write_mtt(dev->dev, &buf->mtt, &buf->buf);
113 if (err)
114 goto err_mtt;
116 return 0;
118 err_mtt:
119 mlx4_mtt_cleanup(dev->dev, &buf->mtt);
121 err_buf:
122 mlx4_buf_free(dev->dev, nent * sizeof(struct mlx4_cqe),
123 &buf->buf);
125 out:
126 return err;
129 static void mlx4_ib_free_cq_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq_buf *buf, int cqe)
131 mlx4_buf_free(dev->dev, (cqe + 1) * sizeof(struct mlx4_cqe), &buf->buf);
134 static int mlx4_ib_get_cq_umem(struct mlx4_ib_dev *dev, struct ib_ucontext *context,
135 struct mlx4_ib_cq_buf *buf, struct ib_umem **umem,
136 u64 buf_addr, int cqe)
138 int err;
140 *umem = ib_umem_get(context, buf_addr, cqe * sizeof (struct mlx4_cqe),
141 IB_ACCESS_LOCAL_WRITE, 1);
142 if (IS_ERR(*umem))
143 return PTR_ERR(*umem);
145 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(*umem),
146 ilog2((*umem)->page_size), &buf->mtt);
147 if (err)
148 goto err_buf;
150 err = mlx4_ib_umem_write_mtt(dev, &buf->mtt, *umem);
151 if (err)
152 goto err_mtt;
154 return 0;
156 err_mtt:
157 mlx4_mtt_cleanup(dev->dev, &buf->mtt);
159 err_buf:
160 ib_umem_release(*umem);
162 return err;
165 struct ib_cq *mlx4_ib_create_cq(struct ib_device *ibdev, int entries, int vector,
166 struct ib_ucontext *context,
167 struct ib_udata *udata)
169 struct mlx4_ib_dev *dev = to_mdev(ibdev);
170 struct mlx4_ib_cq *cq;
171 struct mlx4_uar *uar;
172 int err;
174 if (entries < 1 || entries > dev->dev->caps.max_cqes)
175 return ERR_PTR(-EINVAL);
177 cq = kmalloc(sizeof *cq, GFP_KERNEL);
178 if (!cq)
179 return ERR_PTR(-ENOMEM);
181 entries = roundup_pow_of_two(entries + 1);
182 cq->ibcq.cqe = entries - 1;
183 mutex_init(&cq->resize_mutex);
184 spin_lock_init(&cq->lock);
185 cq->resize_buf = NULL;
186 cq->resize_umem = NULL;
188 if (context) {
189 struct mlx4_ib_create_cq ucmd;
191 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
192 err = -EFAULT;
193 goto err_cq;
196 err = mlx4_ib_get_cq_umem(dev, context, &cq->buf, &cq->umem,
197 ucmd.buf_addr, entries);
198 if (err)
199 goto err_cq;
201 err = mlx4_ib_db_map_user(to_mucontext(context), ucmd.db_addr,
202 &cq->db);
203 if (err)
204 goto err_mtt;
206 uar = &to_mucontext(context)->uar;
207 } else {
208 err = mlx4_db_alloc(dev->dev, &cq->db, 1);
209 if (err)
210 goto err_cq;
212 cq->mcq.set_ci_db = cq->db.db;
213 cq->mcq.arm_db = cq->db.db + 1;
214 *cq->mcq.set_ci_db = 0;
215 *cq->mcq.arm_db = 0;
217 err = mlx4_ib_alloc_cq_buf(dev, &cq->buf, entries);
218 if (err)
219 goto err_db;
221 uar = &dev->priv_uar;
224 err = mlx4_cq_alloc(dev->dev, entries, &cq->buf.mtt, uar,
225 cq->db.dma, &cq->mcq, vector, 0);
226 if (err)
227 goto err_dbmap;
229 cq->mcq.comp = mlx4_ib_cq_comp;
230 cq->mcq.event = mlx4_ib_cq_event;
232 if (context)
233 if (ib_copy_to_udata(udata, &cq->mcq.cqn, sizeof (__u32))) {
234 err = -EFAULT;
235 goto err_dbmap;
238 return &cq->ibcq;
240 err_dbmap:
241 if (context)
242 mlx4_ib_db_unmap_user(to_mucontext(context), &cq->db);
244 err_mtt:
245 mlx4_mtt_cleanup(dev->dev, &cq->buf.mtt);
247 if (context)
248 ib_umem_release(cq->umem);
249 else
250 mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
252 err_db:
253 if (!context)
254 mlx4_db_free(dev->dev, &cq->db);
256 err_cq:
257 kfree(cq);
259 return ERR_PTR(err);
262 static int mlx4_alloc_resize_buf(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
263 int entries)
265 int err;
267 if (cq->resize_buf)
268 return -EBUSY;
270 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
271 if (!cq->resize_buf)
272 return -ENOMEM;
274 err = mlx4_ib_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
275 if (err) {
276 kfree(cq->resize_buf);
277 cq->resize_buf = NULL;
278 return err;
281 cq->resize_buf->cqe = entries - 1;
283 return 0;
286 static int mlx4_alloc_resize_umem(struct mlx4_ib_dev *dev, struct mlx4_ib_cq *cq,
287 int entries, struct ib_udata *udata)
289 struct mlx4_ib_resize_cq ucmd;
290 int err;
292 if (cq->resize_umem)
293 return -EBUSY;
295 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
296 return -EFAULT;
298 cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
299 if (!cq->resize_buf)
300 return -ENOMEM;
302 err = mlx4_ib_get_cq_umem(dev, cq->umem->context, &cq->resize_buf->buf,
303 &cq->resize_umem, ucmd.buf_addr, entries);
304 if (err) {
305 kfree(cq->resize_buf);
306 cq->resize_buf = NULL;
307 return err;
310 cq->resize_buf->cqe = entries - 1;
312 return 0;
315 static int mlx4_ib_get_outstanding_cqes(struct mlx4_ib_cq *cq)
317 u32 i;
319 i = cq->mcq.cons_index;
320 while (get_sw_cqe(cq, i & cq->ibcq.cqe))
321 ++i;
323 return i - cq->mcq.cons_index;
326 static void mlx4_ib_cq_resize_copy_cqes(struct mlx4_ib_cq *cq)
328 struct mlx4_cqe *cqe, *new_cqe;
329 int i;
331 i = cq->mcq.cons_index;
332 cqe = get_cqe(cq, i & cq->ibcq.cqe);
333 while ((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) != MLX4_CQE_OPCODE_RESIZE) {
334 new_cqe = get_cqe_from_buf(&cq->resize_buf->buf,
335 (i + 1) & cq->resize_buf->cqe);
336 memcpy(new_cqe, get_cqe(cq, i & cq->ibcq.cqe), sizeof(struct mlx4_cqe));
337 new_cqe->owner_sr_opcode = (cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK) |
338 (((i + 1) & (cq->resize_buf->cqe + 1)) ? MLX4_CQE_OWNER_MASK : 0);
339 cqe = get_cqe(cq, ++i & cq->ibcq.cqe);
341 ++cq->mcq.cons_index;
344 int mlx4_ib_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
346 struct mlx4_ib_dev *dev = to_mdev(ibcq->device);
347 struct mlx4_ib_cq *cq = to_mcq(ibcq);
348 struct mlx4_mtt mtt;
349 int outst_cqe;
350 int err;
352 mutex_lock(&cq->resize_mutex);
354 if (entries < 1 || entries > dev->dev->caps.max_cqes) {
355 err = -EINVAL;
356 goto out;
359 entries = roundup_pow_of_two(entries + 1);
360 if (entries == ibcq->cqe + 1) {
361 err = 0;
362 goto out;
365 if (ibcq->uobject) {
366 err = mlx4_alloc_resize_umem(dev, cq, entries, udata);
367 if (err)
368 goto out;
369 } else {
370 /* Can't be smaller than the number of outstanding CQEs */
371 outst_cqe = mlx4_ib_get_outstanding_cqes(cq);
372 if (entries < outst_cqe + 1) {
373 err = 0;
374 goto out;
377 err = mlx4_alloc_resize_buf(dev, cq, entries);
378 if (err)
379 goto out;
382 mtt = cq->buf.mtt;
384 err = mlx4_cq_resize(dev->dev, &cq->mcq, entries, &cq->resize_buf->buf.mtt);
385 if (err)
386 goto err_buf;
388 mlx4_mtt_cleanup(dev->dev, &mtt);
389 if (ibcq->uobject) {
390 cq->buf = cq->resize_buf->buf;
391 cq->ibcq.cqe = cq->resize_buf->cqe;
392 ib_umem_release(cq->umem);
393 cq->umem = cq->resize_umem;
395 kfree(cq->resize_buf);
396 cq->resize_buf = NULL;
397 cq->resize_umem = NULL;
398 } else {
399 spin_lock_irq(&cq->lock);
400 if (cq->resize_buf) {
401 mlx4_ib_cq_resize_copy_cqes(cq);
402 mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
403 cq->buf = cq->resize_buf->buf;
404 cq->ibcq.cqe = cq->resize_buf->cqe;
406 kfree(cq->resize_buf);
407 cq->resize_buf = NULL;
409 spin_unlock_irq(&cq->lock);
412 goto out;
414 err_buf:
415 mlx4_mtt_cleanup(dev->dev, &cq->resize_buf->buf.mtt);
416 if (!ibcq->uobject)
417 mlx4_ib_free_cq_buf(dev, &cq->resize_buf->buf,
418 cq->resize_buf->cqe);
420 kfree(cq->resize_buf);
421 cq->resize_buf = NULL;
423 if (cq->resize_umem) {
424 ib_umem_release(cq->resize_umem);
425 cq->resize_umem = NULL;
428 out:
429 mutex_unlock(&cq->resize_mutex);
430 return err;
433 int mlx4_ib_destroy_cq(struct ib_cq *cq)
435 struct mlx4_ib_dev *dev = to_mdev(cq->device);
436 struct mlx4_ib_cq *mcq = to_mcq(cq);
438 mlx4_cq_free(dev->dev, &mcq->mcq);
439 mlx4_mtt_cleanup(dev->dev, &mcq->buf.mtt);
441 if (cq->uobject) {
442 mlx4_ib_db_unmap_user(to_mucontext(cq->uobject->context), &mcq->db);
443 ib_umem_release(mcq->umem);
444 } else {
445 mlx4_ib_free_cq_buf(dev, &mcq->buf, cq->cqe);
446 mlx4_db_free(dev->dev, &mcq->db);
449 kfree(mcq);
451 return 0;
454 static void dump_cqe(void *cqe)
456 __be32 *buf = cqe;
458 printk(KERN_DEBUG "CQE contents %08x %08x %08x %08x %08x %08x %08x %08x\n",
459 be32_to_cpu(buf[0]), be32_to_cpu(buf[1]), be32_to_cpu(buf[2]),
460 be32_to_cpu(buf[3]), be32_to_cpu(buf[4]), be32_to_cpu(buf[5]),
461 be32_to_cpu(buf[6]), be32_to_cpu(buf[7]));
464 static void mlx4_ib_handle_error_cqe(struct mlx4_err_cqe *cqe,
465 struct ib_wc *wc)
467 if (cqe->syndrome == MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR) {
468 printk(KERN_DEBUG "local QP operation err "
469 "(QPN %06x, WQE index %x, vendor syndrome %02x, "
470 "opcode = %02x)\n",
471 be32_to_cpu(cqe->my_qpn), be16_to_cpu(cqe->wqe_index),
472 cqe->vendor_err_syndrome,
473 cqe->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
474 dump_cqe(cqe);
477 switch (cqe->syndrome) {
478 case MLX4_CQE_SYNDROME_LOCAL_LENGTH_ERR:
479 wc->status = IB_WC_LOC_LEN_ERR;
480 break;
481 case MLX4_CQE_SYNDROME_LOCAL_QP_OP_ERR:
482 wc->status = IB_WC_LOC_QP_OP_ERR;
483 break;
484 case MLX4_CQE_SYNDROME_LOCAL_PROT_ERR:
485 wc->status = IB_WC_LOC_PROT_ERR;
486 break;
487 case MLX4_CQE_SYNDROME_WR_FLUSH_ERR:
488 wc->status = IB_WC_WR_FLUSH_ERR;
489 break;
490 case MLX4_CQE_SYNDROME_MW_BIND_ERR:
491 wc->status = IB_WC_MW_BIND_ERR;
492 break;
493 case MLX4_CQE_SYNDROME_BAD_RESP_ERR:
494 wc->status = IB_WC_BAD_RESP_ERR;
495 break;
496 case MLX4_CQE_SYNDROME_LOCAL_ACCESS_ERR:
497 wc->status = IB_WC_LOC_ACCESS_ERR;
498 break;
499 case MLX4_CQE_SYNDROME_REMOTE_INVAL_REQ_ERR:
500 wc->status = IB_WC_REM_INV_REQ_ERR;
501 break;
502 case MLX4_CQE_SYNDROME_REMOTE_ACCESS_ERR:
503 wc->status = IB_WC_REM_ACCESS_ERR;
504 break;
505 case MLX4_CQE_SYNDROME_REMOTE_OP_ERR:
506 wc->status = IB_WC_REM_OP_ERR;
507 break;
508 case MLX4_CQE_SYNDROME_TRANSPORT_RETRY_EXC_ERR:
509 wc->status = IB_WC_RETRY_EXC_ERR;
510 break;
511 case MLX4_CQE_SYNDROME_RNR_RETRY_EXC_ERR:
512 wc->status = IB_WC_RNR_RETRY_EXC_ERR;
513 break;
514 case MLX4_CQE_SYNDROME_REMOTE_ABORTED_ERR:
515 wc->status = IB_WC_REM_ABORT_ERR;
516 break;
517 default:
518 wc->status = IB_WC_GENERAL_ERR;
519 break;
522 wc->vendor_err = cqe->vendor_err_syndrome;
525 static int mlx4_ib_ipoib_csum_ok(__be16 status, __be16 checksum)
527 return ((status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
528 MLX4_CQE_STATUS_IPV4F |
529 MLX4_CQE_STATUS_IPV4OPT |
530 MLX4_CQE_STATUS_IPV6 |
531 MLX4_CQE_STATUS_IPOK)) ==
532 cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
533 MLX4_CQE_STATUS_IPOK)) &&
534 (status & cpu_to_be16(MLX4_CQE_STATUS_UDP |
535 MLX4_CQE_STATUS_TCP)) &&
536 checksum == cpu_to_be16(0xffff);
539 static int mlx4_ib_poll_one(struct mlx4_ib_cq *cq,
540 struct mlx4_ib_qp **cur_qp,
541 struct ib_wc *wc)
543 struct mlx4_cqe *cqe;
544 struct mlx4_qp *mqp;
545 struct mlx4_ib_wq *wq;
546 struct mlx4_ib_srq *srq;
547 int is_send;
548 int is_error;
549 u32 g_mlpath_rqpn;
550 u16 wqe_ctr;
552 repoll:
553 cqe = next_cqe_sw(cq);
554 if (!cqe)
555 return -EAGAIN;
557 ++cq->mcq.cons_index;
560 * Make sure we read CQ entry contents after we've checked the
561 * ownership bit.
563 rmb();
565 is_send = cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK;
566 is_error = (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
567 MLX4_CQE_OPCODE_ERROR;
569 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_OPCODE_NOP &&
570 is_send)) {
571 printk(KERN_WARNING "Completion for NOP opcode detected!\n");
572 return -EINVAL;
575 /* Resize CQ in progress */
576 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) == MLX4_CQE_OPCODE_RESIZE)) {
577 if (cq->resize_buf) {
578 struct mlx4_ib_dev *dev = to_mdev(cq->ibcq.device);
580 mlx4_ib_free_cq_buf(dev, &cq->buf, cq->ibcq.cqe);
581 cq->buf = cq->resize_buf->buf;
582 cq->ibcq.cqe = cq->resize_buf->cqe;
584 kfree(cq->resize_buf);
585 cq->resize_buf = NULL;
588 goto repoll;
591 if (!*cur_qp ||
592 (be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) != (*cur_qp)->mqp.qpn) {
594 * We do not have to take the QP table lock here,
595 * because CQs will be locked while QPs are removed
596 * from the table.
598 mqp = __mlx4_qp_lookup(to_mdev(cq->ibcq.device)->dev,
599 be32_to_cpu(cqe->vlan_my_qpn));
600 if (unlikely(!mqp)) {
601 printk(KERN_WARNING "CQ %06x with entry for unknown QPN %06x\n",
602 cq->mcq.cqn, be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK);
603 return -EINVAL;
606 *cur_qp = to_mibqp(mqp);
609 wc->qp = &(*cur_qp)->ibqp;
611 if (is_send) {
612 wq = &(*cur_qp)->sq;
613 if (!(*cur_qp)->sq_signal_bits) {
614 wqe_ctr = be16_to_cpu(cqe->wqe_index);
615 wq->tail += (u16) (wqe_ctr - (u16) wq->tail);
617 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
618 ++wq->tail;
619 } else if ((*cur_qp)->ibqp.srq) {
620 srq = to_msrq((*cur_qp)->ibqp.srq);
621 wqe_ctr = be16_to_cpu(cqe->wqe_index);
622 wc->wr_id = srq->wrid[wqe_ctr];
623 mlx4_ib_free_srq_wqe(srq, wqe_ctr);
624 } else {
625 wq = &(*cur_qp)->rq;
626 wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
627 ++wq->tail;
630 if (unlikely(is_error)) {
631 mlx4_ib_handle_error_cqe((struct mlx4_err_cqe *) cqe, wc);
632 return 0;
635 wc->status = IB_WC_SUCCESS;
637 if (is_send) {
638 wc->wc_flags = 0;
639 switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
640 case MLX4_OPCODE_RDMA_WRITE_IMM:
641 wc->wc_flags |= IB_WC_WITH_IMM;
642 case MLX4_OPCODE_RDMA_WRITE:
643 wc->opcode = IB_WC_RDMA_WRITE;
644 break;
645 case MLX4_OPCODE_SEND_IMM:
646 wc->wc_flags |= IB_WC_WITH_IMM;
647 case MLX4_OPCODE_SEND:
648 case MLX4_OPCODE_SEND_INVAL:
649 wc->opcode = IB_WC_SEND;
650 break;
651 case MLX4_OPCODE_RDMA_READ:
652 wc->opcode = IB_WC_RDMA_READ;
653 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
654 break;
655 case MLX4_OPCODE_ATOMIC_CS:
656 wc->opcode = IB_WC_COMP_SWAP;
657 wc->byte_len = 8;
658 break;
659 case MLX4_OPCODE_ATOMIC_FA:
660 wc->opcode = IB_WC_FETCH_ADD;
661 wc->byte_len = 8;
662 break;
663 case MLX4_OPCODE_BIND_MW:
664 wc->opcode = IB_WC_BIND_MW;
665 break;
666 case MLX4_OPCODE_LSO:
667 wc->opcode = IB_WC_LSO;
668 break;
669 case MLX4_OPCODE_FMR:
670 wc->opcode = IB_WC_FAST_REG_MR;
671 break;
672 case MLX4_OPCODE_LOCAL_INVAL:
673 wc->opcode = IB_WC_LOCAL_INV;
674 break;
676 } else {
677 wc->byte_len = be32_to_cpu(cqe->byte_cnt);
679 switch (cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) {
680 case MLX4_RECV_OPCODE_RDMA_WRITE_IMM:
681 wc->opcode = IB_WC_RECV_RDMA_WITH_IMM;
682 wc->wc_flags = IB_WC_WITH_IMM;
683 wc->ex.imm_data = cqe->immed_rss_invalid;
684 break;
685 case MLX4_RECV_OPCODE_SEND_INVAL:
686 wc->opcode = IB_WC_RECV;
687 wc->wc_flags = IB_WC_WITH_INVALIDATE;
688 wc->ex.invalidate_rkey = be32_to_cpu(cqe->immed_rss_invalid);
689 break;
690 case MLX4_RECV_OPCODE_SEND:
691 wc->opcode = IB_WC_RECV;
692 wc->wc_flags = 0;
693 break;
694 case MLX4_RECV_OPCODE_SEND_IMM:
695 wc->opcode = IB_WC_RECV;
696 wc->wc_flags = IB_WC_WITH_IMM;
697 wc->ex.imm_data = cqe->immed_rss_invalid;
698 break;
701 wc->slid = be16_to_cpu(cqe->rlid);
702 wc->sl = be16_to_cpu(cqe->sl_vid) >> 12;
703 g_mlpath_rqpn = be32_to_cpu(cqe->g_mlpath_rqpn);
704 wc->src_qp = g_mlpath_rqpn & 0xffffff;
705 wc->dlid_path_bits = (g_mlpath_rqpn >> 24) & 0x7f;
706 wc->wc_flags |= g_mlpath_rqpn & 0x80000000 ? IB_WC_GRH : 0;
707 wc->pkey_index = be32_to_cpu(cqe->immed_rss_invalid) & 0x7f;
708 wc->csum_ok = mlx4_ib_ipoib_csum_ok(cqe->status, cqe->checksum);
711 return 0;
714 int mlx4_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
716 struct mlx4_ib_cq *cq = to_mcq(ibcq);
717 struct mlx4_ib_qp *cur_qp = NULL;
718 unsigned long flags;
719 int npolled;
720 int err = 0;
722 spin_lock_irqsave(&cq->lock, flags);
724 for (npolled = 0; npolled < num_entries; ++npolled) {
725 err = mlx4_ib_poll_one(cq, &cur_qp, wc + npolled);
726 if (err)
727 break;
730 if (npolled)
731 mlx4_cq_set_ci(&cq->mcq);
733 spin_unlock_irqrestore(&cq->lock, flags);
735 if (err == 0 || err == -EAGAIN)
736 return npolled;
737 else
738 return err;
741 int mlx4_ib_arm_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
743 mlx4_cq_arm(&to_mcq(ibcq)->mcq,
744 (flags & IB_CQ_SOLICITED_MASK) == IB_CQ_SOLICITED ?
745 MLX4_CQ_DB_REQ_NOT_SOL : MLX4_CQ_DB_REQ_NOT,
746 to_mdev(ibcq->device)->uar_map,
747 MLX4_GET_DOORBELL_LOCK(&to_mdev(ibcq->device)->uar_lock));
749 return 0;
752 void __mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
754 u32 prod_index;
755 int nfreed = 0;
756 struct mlx4_cqe *cqe, *dest;
757 u8 owner_bit;
760 * First we need to find the current producer index, so we
761 * know where to start cleaning from. It doesn't matter if HW
762 * adds new entries after this loop -- the QP we're worried
763 * about is already in RESET, so the new entries won't come
764 * from our QP and therefore don't need to be checked.
766 for (prod_index = cq->mcq.cons_index; get_sw_cqe(cq, prod_index); ++prod_index)
767 if (prod_index == cq->mcq.cons_index + cq->ibcq.cqe)
768 break;
771 * Now sweep backwards through the CQ, removing CQ entries
772 * that match our QP by copying older entries on top of them.
774 while ((int) --prod_index - (int) cq->mcq.cons_index >= 0) {
775 cqe = get_cqe(cq, prod_index & cq->ibcq.cqe);
776 if ((be32_to_cpu(cqe->vlan_my_qpn) & MLX4_CQE_QPN_MASK) == qpn) {
777 if (srq && !(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK))
778 mlx4_ib_free_srq_wqe(srq, be16_to_cpu(cqe->wqe_index));
779 ++nfreed;
780 } else if (nfreed) {
781 dest = get_cqe(cq, (prod_index + nfreed) & cq->ibcq.cqe);
782 owner_bit = dest->owner_sr_opcode & MLX4_CQE_OWNER_MASK;
783 memcpy(dest, cqe, sizeof *cqe);
784 dest->owner_sr_opcode = owner_bit |
785 (dest->owner_sr_opcode & ~MLX4_CQE_OWNER_MASK);
789 if (nfreed) {
790 cq->mcq.cons_index += nfreed;
792 * Make sure update of buffer contents is done before
793 * updating consumer index.
795 wmb();
796 mlx4_cq_set_ci(&cq->mcq);
800 void mlx4_ib_cq_clean(struct mlx4_ib_cq *cq, u32 qpn, struct mlx4_ib_srq *srq)
802 spin_lock_irq(&cq->lock);
803 __mlx4_ib_cq_clean(cq, qpn, srq);
804 spin_unlock_irq(&cq->lock);